Conjecture—Morphology should be co-designed with learning

Investigate how physical morphology and learning methodology should be co-designed rather than treated in isolation to improve learning, robustness, and capability in embodied agents.

Background

Robot morphology—sensors, actuators, dynamics—induces strong inductive biases affecting what can be learned and how reliably.

The authors conjecture that effective learning methods should complement and guide physical design choices, arguing against treating morphology and learning separately.

References

We conjecture that physical design is significantly complemented by effective learning methodology and should not be treated in isolation. We discuss this conjecture in section \ref{sec:hardware}.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 1 (Introduction: Guiding Questions)

There exists a complex and as-yet-unresolved interplay among environmental complexity, agent morphology, and learning ability.

— Towards general embodied intelligence: integrating large language models, knowledge bases, and reasoning capabilities to build the next generation of AI agents  (2608.19794 - Yuan et al., 20 Aug 2026) in Section 2, subsection “Analysis of Difficulties of Embodied Intelligence”